The 180-Degree Turn
At the end of 2025, the compute market was still worried about unsold inventory.
By early 2026, the winds had completely shifted. Compute became scarce, prices began rising. Kimi's API revenue during just a few days of the Lunar New Year equaled its entire previous year.
The reason isn't complicated: how we use AI has changed. From conversations to agents, compute consumption multiplied 100-fold in just two to three months. Data shows that in Agent scenarios, per-user token consumption is 10 to 50 times that of ordinary chat, with some scenarios seeing increases of up to 370x.
Cloud compute is being utilized quite efficiently—tech giants are scrambling for GPUs, startups are queuing up, and resources aren't going to waste.
But edge compute is another story.
The Edge Dilemma
AI PCs, edge inference servers, various AI accelerators—these devices are evolving rapidly, and prices are dropping fast. NVIDIA's DGX Spark costs around 30,000 RMB, while AMD's Ryzen AI Max+ 395 is even under 20,000. The compute is substantial, capable of inferencing fairly large models.
But are they being used well?
From what I see, the answer is: not well.
The problem lies in Total Cost of Ownership (TCO). The devices themselves have become cheaper, but the proportion of other cost items has been violently amplified.
Operations and Maintenance Costs
A 200,000 RMB server fails twice a year, costing 1,000 RMB each time for technician visits—2,000 RMB annually. That's 1% of the device cost, negligible.
Switch to a 20,000 RMB edge device. The failure rate won't be lower—it might even be higher, because edge environments are more complex. Still 2,000 RMB in O&M costs.
But now that's 10% of the cost. From 1% to 10%—a 10x amplification.
Reality is worse. On-site maintenance for edge devices can cost 500 RMB per visit per device, and with dispersed deployments, inspection efficiency is low.
Cold Start Costs
A purchased device is just a piece of iron.
How do you turn it into real AI compute? How do you extract its performance?
It's harder than imagined. Hardware evolves daily, inference engines vary widely, models change too fast, heterogeneous architectures are headaches. ARM, x86, RISC-V coexist, firmware versions are fragmented. One company with 2,000 edge devices saw fault diagnosis time triple due to inconsistent firmware versions.
To use these devices well, you need someone who understands hardware, inference engines, models, and applications. A 20,000 RMB monthly salary for such a person isn't unreasonable.
But the device itself only costs 20,000 RMB.
That's equivalent to paying a 100% premium just to utilize a device. This doesn't make business sense.
Ongoing Usage Costs
Getting the device running isn't the end—it's the beginning.
In 2026, the pace of updates from leading model providers is dizzying. Every month or two, a major new model drops. Google Gemini iterated from 2.5 to 3 Flash within a year, Alibaba's Qwen3 series launched in rapid succession, DeepSeek and Kimi took turns entering the spotlight.
If your edge device is still running a model from a year ago, its value might be down to 10%.
Keeping up with this pace means continuous upgrading, adaptation, testing, and deployment. These investments are calculated per headcount—they don't get discounted just because the device is cheaper.
Do People Really Want to Be Involved?
Traditional software's solution approach is: design a system for humans to operate.
This approach has an implicit assumption: that humans are willing to participate.
But in the edge AI scenario, does this assumption hold?
Think about electricity. People use it but don't participate in its production or management. Electricity is infrastructure—people just consume it. No one feels the need to "participate in the production process of electricity."
Edge AI devices should be the same.
They are infrastructure, not tools. The ideal state is devices that self-run, self-heal, self-upgrade, and never fail.
If humans don't have to participate, all the better.
This means a shift in design philosophy: let devices manage themselves, keep humans outside the loop, software isn't a tool for humans but the "nervous system" of the device. The goal isn't to reduce human burden—it's to eliminate human participation.
How to Get There?
I see several trends we can leverage.
First, AI Agent capabilities themselves are rapidly strengthening. If your product doesn't let Agents participate in core workflows, it's out of touch with the era. Agents can complete O&M diagnostics, fault repair, performance tuning, and model upgrades without human intervention—automating the work of that "20,000 RMB monthly salary person." This isn't future tense; Agent capabilities in 2026 are already sufficient.
Second, open-source community collaboration. Edge AI faces a four-dimensional optimization problem: hardware devices × inference engines × models × applications. Each dimension is changing rapidly; no single company can keep up. But with Agent adoption, code-repository-based collaboration becomes easier. Products need to consider how to get more people and Agents involved to collectively address the growth in complexity.
Finally, internet infrastructure. Edge devices are characterized by fragmentation—AI PCs and edge servers distributed across offices, factories, stores, and homes, unlike data centers where thousands are centrally managed in one room. To extract value from these fragmented devices, you must use the internet. Remote monitoring, OTA upgrades, cloud collaboration—these aren't optional.
Final Thoughts
The TCO problem of edge AI is essentially a management philosophy problem.
Continuing to design products with a "human-centered" approach will keep amplifying O&M costs, keeping cold start costs high, and letting ongoing usage costs continue to devour value.
The solution direction is to let AI participate. Use Agents to automate operations, use the community to distribute complexity, use the internet to connect fragmented devices.
This is what we're working on.